T9a: Parser reads real rules: . as subject, ../field alias, function forms #14

Parent: T9 (#9) · Importance: Must have · Size: M · Depends on: nothing. Unblocks 9e and 9g.

Why. The shared parser only recognises comparisons whose subject is ${field}. Real configs write
validation rules against the answer itself as . (. >= 0 and . <= 20), and real relevants use
relative paths (../lmp_approx). Across 777 constraint rules in seven configs the builders open
3. Nothing in the UI can improve until the parser reads what real forms contain. This slice is
shared-only and carries the round-trip risk, so it ships alone, before any UI depends on it.

Scope.

• relevantParser.ts: accept . as a comparison subject. Values may be a number, a quoted string,
  ${field}, ../field, today(), now(), today() ± N, date('…').
• Accept ../field as an alias of ${field} for subject and value. ⚠ Re-emit exactly as written;
  never rewrite ../field to ${field} or back.
• Recognise the function forms on . that real configs use: string-length(.), regex(., '…'),
  selected(., '…'), count-selected(.), int(.), int(format-date(., '%Y')),
  max(coalesce(${a}, 0), …), add-date(today(), …), difference-in-months(., today()),
  decimal-date-time(.).
• Recognise true, true() and 1 as an always-true rule kind so the UI can label it. ⚠ Never
  rewritten.
• ⚠ Additive only. New rule kinds or new entry points; parseRelevant / parseRelevantGrouped
  return shapes seen by the five existing consumers do not change.
• The parse → serialize self-check stays authoritative: anything that does not re-serialize
  byte-identical is raw text. .<=100 and . <= 100 both round-trip unchanged.
• node --test round-trip case per new form in shared/src/xlsform/, each calling the serializer
  and starting from a non-canonical fixture.
• Convert the ../field hostile fixtures from { todo: true } to live for the byte-stability half;
  only the "opens as clauses" half stays todo until 9e / 9f.
• QA: measure the ../ vs ${} share across the seven configs and record it in the plan folder.

Not in scope. Any UI. The preset recogniser (9e). The function catalogue (separate issue).

Acceptance.

1. scripts/corpus-sweep.mjs shows zero drift on all seven configs after the change.
2. A test over the 777-rule corpus reports how many constraint rules now parse structurally; the
   number is recorded in the README of the plan folder (target after 9e: about 705 of 726).
3. Every new round-trip test starts from a non-canonical fixture and asserts
   serialize(parse(x)) === x.
4. pnpm typecheck is clean with no change to any consumer of parseRelevant, which proves the
   widening is additive.

Validate.

pnpm --filter @cht-ui/shared build && pnpm --filter @cht-ui/shared test
pnpm typecheck && pnpm lint
node scripts/smoke-parser.mjs <config>/forms/app/pregnancy.xlsx   # Round-trip stable: YES
node scripts/corpus-sweep.mjs

Plan: docs/plans/9_complex_logic_calculation_relevant_constraint/README.md · Analysis: validation-rules-v2.md §"Proposed support" 1, §"Test plan". Line anchors against bff69cb: relevantParser.ts:370 is the ${}-only comparison regex.